Competitive attraction in neural networks with sign-constrained weights
K. Y. Michael Wong, Colin K. Campbell · Journal of Physics A Mathematical and General · 1992
The presence of many attractors in neural networks give rise to interesting competitive phenomena. The authors consider dilute recurrent (or attractor) neural networks with sign-constrained weights and storing uncorrelated patterns with maximal stability. The dynamics of these networks is governed by the competitive effects of retrieval, nonretrieval and uniform (i.e. ferromagnetic) attractors, which result in basin encroaching, shrinking, splitting and wedging. They have found the parameter regions in which each of these attractors exist. The basins of attraction of the uniform attractors enlarge at the expense of the other attractors even when the weight signs are slightly imbalanced, but can be compensated by the introduction of a dynamical threshold.